betaFeedback

WorksMathematics

The Exact Central-Index Profile of Groups with Bounded Noncommutativity

Abstract

For a finite group GG, let ω(G)ω(G) be the largest size of a pairwise noncommuting set and put z(G)=[G:Z(G)]z(G)=[G:\mathrm Z(G)]. We determine the maximum possible central index under ω(G)≤kω(G)\le k for every integer k≥1k\ge1, giving an exact answer to Neumann's central-index question and sharpening Pyber's exponential upper bound. The maximum is 11 for k=1,2k=1,2, (k−1)2(k-1)^2 for 3≤k≤63\le k\le6, 8181 for k=7,8k=7,8, and 4⌊(k−1)/2⌋4^{\lfloor(k-1)/2\rfloor} for k≥9k\ge9. In the last range, with r=⌊(k−1)/2⌋r=\lfloor(k-1)/2\rfloor, equality holds exactly when G′≅C2G'\cong C_2 and G/Z(G)≅F22rG/\mathrm Z(G)\cong\mathbb F_2^{2r}. We also obtain the sharp universal bound z(G)≤(81/64)2ω(G)−1z(G)\le(81/64)2^{ω(G)-1}. All maximizers are classified up to isoclinism. The results extend to arbitrary groups with finite clique number through finite isoclinic representatives.

Introduction

The size of a pairwise noncommuting set measures how many elements of a group can be mutually incompatible with commutativity. For a finite group GG, let ω(G)\omega(G) be the largest size of such a set. We ask how strongly a bound on ω(G)\omega(G) constrains the index of the centre. The order of GG itself cannot be bounded in this way: taking a direct product with an arbitrarily large abelian group changes neither ω(G)\omega(G) nor the central index. The natural extremal quantities are therefore

z(G)=[G:Z(G)],ζ(k)=sup⁡{z(G):ω(G)≤k}.z(G)=[G:Z(G)],\qquad\zeta(k)=\sup\{z(G):\omega(G)\le k\}.

Here the supremum is over finite groups. In graph-theoretic language, join two elements of GG when they do not commute; then ω(G)\omega(G) is the clique number of this noncommuting graph. Thus the problem relates a combinatorial restriction to a structural invariant of the group.

The origin is a question posed by Erdős in 1975: if a group has no infinite pairwise noncommuting subset, must the sizes of its finite pairwise noncommuting subsets be bounded? Neumann answered this affirmatively in 1976 by proving the stronger statement that such a group has finite central quotient [20 Theorem 6]. We first solve the finite extremal problem and then extend the sharp bounds to arbitrary groups by isoclinism, as explained in Corollary 1.4 below. Neumann also explicitly asked how the central index can be bounded in terms of the largest noncommuting set. His proof gave log⁡z(G)=O(ω(G)2)\log z(G)=O(\omega(G)^2), and he suggested a possible quadratic bound for z(G)z(G) itself [20 Section 4].

Pyber resolved the scale of this quantitative question in 1987: there is an absolute constant cc such that z(G)≤cω(G)z(G)\le c^{\omega(G)} for every finite group [21]. In particular, ζ(k)\zeta(k) is finite. The extraspecial 22-groups show that exponential growth is necessary; their precise parameters are recalled below.

The remaining extremal questions are to determine the optimal exponential base, the exact maximum at each finite cap, and the groups attaining it. These are distinct from the related questions about large abelian subgroups and the number of commuting tuples studied by Erdős and Straus [12].

Several parts of the literature supply information about the models and local estimates relevant to these questions. Isaacs’s formula, reported by Bertram, gives the clique numbers of extraspecial 2-groups [5 p. 40]; Chin studies noncommuting sets in extraspecial pp-groups [10]. Further results treat groups with derived subgroup of prime order and metacyclic pp-groups [18, 14]. For scalar commutator maps, noncommuting sets correspond to sets of pairwise nonorthogonal points in symplectic geometry. Results on partial ovoids therefore provide sharp information in some small odd-characteristic cases [7, 9]. Covering theorems provide a complementary tool: the centralizers of the elements of a maximum clique cover the group, so bounds for subgroup and coset covers can control the central index [19, 22, 8].

The present article determines ζ(k)\zeta(k) and classifies all its maximizers up to isoclinism for every k≥1k \ge1. The eventual extremizers are precisely the extraspecial binary isoclinism classes, but small parameters also require quinary and ternary models. The proof combines the geometry of the nilpotent cases with a centralizer descent that excludes larger values in arbitrary finite groups. We first state the exact profile, then describe its models, its extension to arbitrary groups, and the structure of the proof.

Theorem 1.1 (The central-index profile). For every integer k≥1k \ge1, the supremum defining ζ(k)\zeta(k) is attained and equals

M(k)={1,k=1,2,(k−1)2,3≤k≤6,81,k=7,8,4⌊(k−1)/2⌋,k≥9.(1)M(k) = \begin{cases} 1, & k = 1,2, \\ (k-1)^2, & 3 \le k \le6, \\ 81, & k = 7,8, \\ 4^{\lfloor(k-1)/2 \rfloor}, & k \ge9. \end{cases} \tag*{(1)}

The three nontrivial ranges have different extremal models. In the Heisenberg group UT3(Fq)\mathrm{UT}_3(\mathbb{F}_q), the commutator on the central quotient Fq2\mathbb{F}_q^2 is the determinant. Pairwise noncommuting elements project to distinct lines, and one may choose one representative of every line. Consequently

z(UT3(Fq))=q2,ω(UT3(Fq))=q+1.z(\mathrm{UT}_3(\mathbb{F}_q)) = q^2,\qquad\omega(\mathrm{UT}_3(\mathbb{F}_q)) = q+1.

Taking q=2,3,4,5q=2,3,4,5 explains the four initial squares. An extraspecial 3-group of order 353^5 instead has z=81z=81 and ω=7\omega=7. Its cliques are partial ovoids of W(3,3)W(3,3), that is, sets of pairwise nonorthogonal projective points: Tallini’s bound q2−q+1q^2-q+1 for odd qq gives the upper bound seven, and two symplectic planes give the matching construction [9 Section 2 and Table 1]; see also [7, 11]. Finally, for an extraspecial 2-group ErE_r of order 21+2r2^{1+2r}, Isaacs’s clique formula, reported by Bertram [5 p. 40], gives

z(Er)=4r,ω(Er)=2r+1.z(E_r)=4^r,\qquad\omega(E_r)=2r+1.

Thus the largest allowed rr gives Lk=4⌊(k−1)/2⌋L_k=4^{\lfloor(k-1)/2 \rfloor} and explains the paired plateaux. These examples prove all lower bounds in Theorem 1.1.

Theorem 1.2 (All extremizers). Let GG be finite with ω(G)≤k\omega(G) \le k.

  1. If k=1k=1 or 22, then GG is abelian and z(G)=1z(G)=1.

  1. If k∈{3,4,6}k \in\{3,4,6\} and p=k−1p=k-1, then z(G)=p2z(G)=p^2 if and only if G′≅CpG' \cong C_p and G/Z(G)≅Fp2G/Z(G) \cong\mathbb{F}_p^2; such groups have ω(G)=k\omega(G)=k and are isoclinic to an extraspecial group of order p3p^3.

  1. If k=5k=5, then z(G)=16z(G)=16 if and only if GG is isoclinic to either an extraspecial group of order 3232 or UT3(F4)\mathrm{UT}_3(\mathbb{F}_4). These are two distinct isoclinism classes, and both have ω(G)=5\omega(G)=5.

(iv) If k=7k = 7 or 88, then z(G)=81z(G) = 81 if and only if G′≅C3G' \cong C_3 and G/Z(G)≅F34G/Z(G) \cong\mathbb{F}_3^4; such groups have ω(G)=7\omega(G) = 7.

(v) If k≥9k \ge9 and r=⌊(k−1)/2⌋r = \lfloor(k - 1)/2 \rfloor, then z(G)=Lkz(G) = L_k if and only if G′≅C2G' \cong C_2 and G/Z(G)≅F22rG/Z(G) \cong\mathbb{F}_2^{2r}; such groups have ω(G)=2r+1\omega(G) = 2r + 1.

In (iv) and (v), the groups form the corresponding extraspecial isoclinism class. Thus all maximizers are identified up to isoclinism for every cap kk.

Corollary 1.3 (Sharp exponential bound). Every finite group satisfies

z(G)≤81642ω(G)−1.z(G) \le\frac{81}{64}2^{\omega(G)-1}.

Equality holds precisely when G′≅C3G' \cong C_3 and G/Z(G)≅F34G/Z(G) \cong\mathbb{F}_3^4. Moreover,

z(G)≤2ω(G)−1if ω(G)≥9,lim⁡k→∞ζ(k)1/k=2.(2)z(G) \le2^{\omega(G)-1} \quad\text{if } \omega(G) \ge9,\qquad\lim_{k\to\infty}\zeta(k)^{1/k}=2. \tag*{(2)}

Corollary 1.4 (Arbitrary groups). Let GG be a group, not necessarily finite, with ω(G)≤k<∞\omega(G) \le k < \infty. Then [G:Z(G)]≤M(k)[G : Z(G)] \le M(k), and all the equality characterizations in Theorem 1.2 remain valid. The sharp bound and its equality case in Corollary 1.3 also remain valid when ω(G)<∞\omega(G) < \infty. For every kk, the bound M(k)M(k) is attained by both finite and infinite groups.

Proof. The abelian case is immediate. Otherwise, Zarrin’s finite-representative theorem [23 Theorem 1.1(2)] gives a finite group HH isoclinic to GG, with ω(H)=ω(G)\omega(H) = \omega(G). Isoclinism also gives H/Z(H)≅G/Z(G)H/Z(H) \cong G/Z(G) and H′≅G′H' \cong G'. Thus z(H)=z(G)z(H) = z(G), and the finite results give every asserted bound and equality characterization. Taking the direct product of a finite maximizer with an infinite abelian group preserves both invariants and proves the last assertion. □

Underlying this established reduction are the finiteness of the central quotient, Schur’s theorem, and Hall’s existence of a stem representative HH, meaning that Z(H)≤H′Z(H) \le H' [17]. Isoclinism preserves the commutator map and hence the clique number [23 Lemma 2.1 and proof of Theorem 1.1]. In particular, the equality statements specify an isoclinism class; they do not assert a direct-product decomposition of an infinite extremizer.

Proof strategy. The proof separates the nilpotent structure from the passage to a general group. § 2 develops one calculus for both: cliques in cosets and commuting products. § 3 proves the Sylow budgets, the complete nilpotent profile and the binary rigidity theorem.

For a centralizer K=CG(L)K = C_G(L), the accounting identity is

z(G)=[G:K]z(K)[Z(K):Z(G)].(3)z(G) = [G : K]z(K)[Z(K) : Z(G)]. \tag*{(3)}

§ 4 introduces the prime cost of losing centrality and then uses Baer–Suzuki to choose a nonnilpotent LL, whose central quotient amplifies cliques of KK four times. Two saturation principles close the finite exceptions: near-minimal ternary losses force divisibility, and affine conjugacy classes force a faithful action on the class.

§ 5 first settles the finite caps and then begins one induction at eighteen. A numerical prime-cost bound pays for the internal centralizer index with the same budget as the central loss. This leaves one centralizer estimate, a finite list of saturation obstructions and one envelope for all larger caps. The extremizer statements then follow from the nilpotent rigidity already proved, and the sharp exponential estimate follows from the exact profile.

Notation. Groups in the proof of the finite results are finite; Corollary 1.4 explicitly removes that hypothesis. For an arbitrary group, ω(G)\omega(G) denotes the supremum of the cardinalities of pairwise noncommuting subsets; whenever it is finite, it is a maximum. We write G′G' for the derived subgroup and CG(X)C_G(X) for the pointwise centralizer of a subset XX, with CG(u,v)=CG(u)∩CG(v)C_G(u,v)=C_G(u)\cap C_G(v). We use [x,y]=x−1y−1xy[x,y]=x^{-1}y^{-1}xy and xg=g−1xgx^g=g^{-1}xg. A clique is a pairwise noncommuting set; maximum means of largest cardinality. We also use ω(X)\omega(X) for a subset or a coset. Two sets are completely adjacent if every element of one fails to commute with every element of the other. An abelian group has clique number one, and no group has clique number two: a noncommuting pair x,yx,y extends to the clique x,y,xyx,y,xy. The integer k≥1k\ge1 is an upper cap; for a fixed ambient group GG, write N=ω(G)N=\omega(G) and Z=Z(G)Z=Z(G). Logarithms are binary unless specified. CmC_m denotes the cyclic group of order mm and Fq\mathbb{F}_q the field with qq elements. We write W(2r−1,q)W(2r-1,q) for the symplectic polar space associated with Fq2r\mathbb{F}_q^{2r}. For a finite abelian or nilpotent group HH, the symbol H(p)H_{(p)} denotes its Sylow pp-subgroup. We use γi(P)\gamma_i(P) for the lower central series and κ=81/64\kappa=81/64 for the sharp constant above. Isoclinism means an identification of the central quotients and derived groups compatible with their commutator maps; it does not require a direct-product decomposition with an extraspecial factor.

Clique constructions

We use three ways to construct cliques: inside a coset, by adjoining an element that moves a central subgroup, and across commuting factors. Only the coset estimate requires a covering argument.

Cliques in a coset

A coset cover is irredundant if none of its members can be deleted. A maximum clique in a coset will produce such a cover by centralizers; the next estimate turns that cover into a lower bound for the clique. For its context see [19, 22, 8, 4]; the binary vector-space case is also covered by [6].

Lemma 2.1 (Coset-covering bound for a pp-group). Let UU be a finite pp-group and U=⋃i=1tHixiU=\bigcup_{i=1}^{t}H_ix_i an irredundant covering by right cosets of subgroups. Then

[U:⋂i=1tHi]≤pt−1.\left[U:\bigcap_{i=1}^{t}H_i\right]\le p^{t-1}.

Proof. We induct on ∣U∣|U|, the case t=1t=1 being immediate. For t>1t>1 all HiH_i are proper. Choose a maximal subgroup M≥H1M\ge H_1; it is normal of index pp. Translate the cover so that H1x1∩M=∅H_1x_1\cap M=\varnothing, and choose an irredundant subcover of MM with indices I1⊆{2,…,t}I_1\subseteq\{2,\ldots,t\}. Put B1=⋂i∈I1HiB_1=\bigcap_{i\in I_1}H_i and D=⋂iHiD=\bigcap_iH_i. Induction gives [M:M∩B1]≤p∣I1∣−1[M:M\cap B_1]\le p^{|I_1|-1}.

Suppose the subgroup intersection for the selected indices is Bj≠DB_j\ne D. Their cosets form a proper subfamily of the original irredundant cover, so choose yy outside their union. The coset BjyB_jy avoids them all. Intersect the remaining cosets with BjyB_jy and translate by y−1y^{-1}; their nonempty intersections are right cosets in BjB_j, with directions Bj∩HiB_j\cap H_i. An irredundant subcover selects new indices Ij+1I_{j+1} and gives

Bj+1=Bj∩⋂i∈Ij+1Hi,[Bj:Bj+1]≤p∣Ij+1∣−1.B_{j+1}=B_j\cap\bigcap_{i\in I_{j+1}}H_i,\qquad[B_j:B_{j+1}]\le p^{|I_{j+1}|-1}.

Induction applies since Bj<B1<UB_j<B_1<U. The nonempty index sets are disjoint, so this process reaches Bs=DB_s=D after finitely many steps, even if some successive subgroups coincide. Since D≤H1≤MD\le H_1\le M,

[U:D]≤p[M:M∩B1][B1:D]≤p1+∑j=1s(∣Ij∣−1)≤pt−1.[U:D]\le p[M:M\cap B_1][B_1:D]\le p^{1+\sum_{j=1}^{s}(|I_j|-1)}\le p^{t-1}.

For the last inequality, s=1s=1 gives ∣I1∣≤t−1|I_1| \le t-1; if s≥2s \ge2, use ∑j∣Ij∣≤t\sum_j |I_j| \le t. No subgroup direction is assumed normal.

Lemma 2.2 (Coset calculus). Let U≤GU \le G, g∈Gg \in G, and let XX be a maximum clique in gUgU. Then

U∩CG(X)=Z(U)∩CU(g).U \cap C_G(X) = Z(U) \cap C_U(g).

The sets CG(x)∩gUC_G(x) \cap gU, x∈Xx \in X, form an irredundant coset cover, and some x∈Xx \in X satisfies [U:CU(x)]≤∣X∣[U : C_U(x)] \le|X|. If UU is a pp-group, then

ω(gU)≥1+log⁡p[U:Z(U)∩CU(g)].(4)\omega(gU) \ge1 + \log_p[U : Z(U) \cap C_U(g)]. \tag*{(4)}

If UU is abelian, the exact formula is

ω(gU)=[U:CU(g)].(5)\omega(gU) = [U : C_U(g)]. \tag*{(5)}

If UU is a nonabelian pp-group, then ω(gU)≥p+1\omega(gU) \ge p+1.

Proof. The right-hand side of the first identity centralizes all of gUgU. Conversely, suppose a∈Ua \in U centralizes XX but not some y∈gUy \in gU. Multiply by aa precisely the vertices of XX commuting with yy. Mutual noncommutation is preserved, and the new clique extends by yy, a contradiction. This proves the identity.

Maximality makes the displayed sets cover gUgU, and each x∈Xx \in X belongs only to its own set, so the cover is irredundant. The part at xx is xCU(x)xC_U(x) and has size ∣CU(x)∣|C_U(x)|; counting gives the averaging assertion. Translation by g−1g^{-1} identifies the parts with left cosets in UU. Inversion turns them into right cosets with the same subgroup directions. The intersection of those directions is Z(U)∩CU(g)Z(U) \cap C_U(g), so Lemma 2.1 proves (4). Finally, when UU is abelian, gaga and gbgb commute exactly when ab−1∈CU(g)ab^{-1} \in C_U(g), proving (5).

For a nonabelian pp-group, a one-part cover would force U=Z(U)∩CU(g)U = Z(U) \cap C_U(g). Otherwise every direction is proper and each part has size at most ∣U∣/p|U|/p. At most pp parts could cover only by partitioning into pp index-pp cosets. Their directions are normal, and cosets of distinct index-pp subgroups intersect, so all directions coincide. Their intersection lies in Z(U)Z(U), making U/Z(U)U/Z(U) cyclic, a contradiction.

Lemma 2.3 (Clique extensions). If A≤Z(H)A \le Z(H) and [A,g]≠1[A,g] \ne1, then

ω(⟨H,g⟩)≥ω(H)+[A:CA(g)].(6)\omega(\langle H,g\rangle) \ge\omega(H) + [A : C_A(g)]. \tag*{(6)}

If [K,L]=1[K,L] = 1, then

ω(KL)≥ω(K)ω(L/Z(L)).(7)\omega(KL) \ge\omega(K)\omega(L/Z(L)). \tag*{(7)}

Proof. The abelian-coset formula gives [A:CA(g)][A : C_A(g)] vertices in gAgA. Choose a0∈A∖CA(g)a_0 \in A \setminus C_A(g). In a maximum clique of HH, multiply by a0a_0 precisely the vertices commuting with gg. Their mutual commutators are preserved, and every resulting vertex is noncommuting with every element of gAgA. This proves (6).

For (7), multiply a maximum clique of KK by lifts of a clique in L/Z(L)L/Z(L). Cross commutators from LL lie outside Z(L)Z(L), whereas K∩L≤Z(L)K \cap L \le Z(L), so they cannot cancel commutators from KK. Products using the same lift remain noncommuting inside KK.

Lemma 2.4 (Small central quotients). If QQ is nonabelian and ω(Q)≤4\omega(Q) \le4, then Q/Z(Q)Q/Z(Q) is C22C_2^2, C32C_3^2, or S3S_3. Thus z(Q)=4z(Q)=4 for ω(Q)=3\omega(Q)=3, and z(Q)≤9z(Q) \le9 for ω(Q)=4\omega(Q)=4.

Proof. For x≠Z(Q)x \ne Z(Q), twisting on ⟨x⟩≤Z(CQ(x))\langle x\rangle\le Z(C_Q(x)) gives ω(CQ(x))≤ω(Q)−2≤2\omega(C_Q(x)) \le\omega(Q)-2 \le2. Every proper centralizer is therefore abelian. Their noncentral parts partition Q∖Z(Q)Q \setminus Z(Q), and their number is s=ω(Q)∈{3,4}s = \omega(Q) \in\{3,4\}. Write their orders modulo Z(Q)Z(Q) as m1≥⋯≥ms≥2m_1 \ge\cdots\ge m_s \ge2 and put q=z(Q)q = z(Q). Then

q=1+∑i(mi−1),m1m2≤q≤m1+(s−1)(m2−1),q = 1 + \sum_i (m_i - 1), \qquad m_1m_2 \le q \le m_1 + (s - 1)(m_2 - 1),

so m1≤s−1m_1 \le s - 1. For s=3s = 3 all orders are two, giving q=4q = 4. For s=4s = 4, the conditions mi∣qm_i \mid q leave only (3,2,2,2)(3,2,2,2) or (3,3,3,3)(3,3,3,3), giving q=6q = 6 or nine. A cyclic central quotient is impossible; the listed isomorphism types follow.

The nilpotent profile and rigidity

The common coset calculus controls every Sylow subgroup through a scalar commutator form and the cliques in its fibres. We prove the prime budgets first, combine them into the nilpotent profile, and then establish the binary gap that will classify the maximizers. We use the elementary fact that if G/Z(G)G/Z(G) is a pp-group, lifting its central series makes GG nilpotent and all Sylow subgroups at other primes central.

The symplectic model

For a symplectic form bb, a clique means a set whose distinct vectors have nonzero bb-pairing. Such sets model noncommuting elements [5, 10, 18]. The hyperplane variant below will also be used in the equality proof.

Lemma 3.1 (Symplectic cliques). Let (V,b)(V,b) be symplectic of dimension 2r>02r > 0 over Fp\mathbb{F}_p. There is a clique of size pr+1p^r + 1, and one can choose such a clique to meet any prescribed hyperplane in exactly one point. Every maximum clique spans VV. For p=2p = 2, the maximum is exactly 2r+12r + 1, and every maximum clique has sum zero. For p=3p = 3, a clique of size at least 4r−14r - 1 exists.

Proof. In a symplectic plane, choose one representative of each of its p+1p + 1 lines. If X,YX,Y are cliques in orthogonal summands and y0∈Yy_0 \in Y, then (X+y0)∪(Y∖{y0})(X + y_0) \cup(Y \setminus\{y_0\}) is a clique of size ∣X∣+∣Y∣−1|X| + |Y| - 1. This proves the lower bound pr+1p^r + 1. For the hyperplane assertion write J=e⊥J = e^\perp, choose b(e,f)=1b(e,f) = 1, and split V=⟨e,f⟩⊥V0V = \langle e,f\rangle\perp V_0. For a clique D0⊂V0D_0 \subset V_0 of size p(r−1)+1p^{(r-1)} + 1, using D0={0}D_0 = \{0\} when r=1r = 1, the set

(D0+f)∪{e+tf:t∈Fp}(D_0 + f) \cup\{e + tf : t \in\mathbb{F}_p\}

is a clique of size pr+1p^r + 1, and only ee belongs to JJ.

If a maximum clique DD does not span, choose 0≠v∈⟨D⟩⊥0 \ne v \in\langle D\rangle^\perp and yy with b(v,y)≠0b(v,y) \ne0. Translate by vv precisely the vertices of DD orthogonal to yy. All mutual pairings are unchanged, and the resulting clique extends by yy, a contradiction.

Over F2\mathbb{F}_2, a clique of size tt has Gram matrix Jt−ItJ_t - I_t, where JtJ_t has all entries one and ItI_t is the identity matrix. The Gram matrix has rank tt for even tt and t−1t - 1 for odd tt. Hence t≤2r+1t \le2r + 1, with equality by the construction above. The sum of a maximum clique is orthogonal to each of its vertices; as these span, the sum is zero. For p=3p = 3, [9 Theorem 3.7] supplies a partial ovoid of W(5,3)W(5,3) of size 32+3+1=133^2 + 3 + 1 = 13. Choosing one nonzero vector on each projective point gives thirteen vertices in dimension six. For r=3u+vr = 3u + v, 0≤v≤20 \le v \le2, orthogonal gluing with vv planes gives 12u+3v+1≥4r−112u + 3v + 1 \ge4r - 1 vertices.

The next lemma enlarges cliques when a scalar form hides a second commutator coordinate. It will be used in the small ternary case and in the binary rigidity proof.

Lemma 3.2 (Affine inflation of a nonscalar form). Let β:V×V→W\beta: V \times V \to W be alternating bilinear over Fp\mathbb{F}_p, with values spanning WW, dim⁡V=2r\dim V = 2r and dim⁡W≥2\dim W \ge2. If some scalarization b=λ∘βb = \lambda\circ\beta is nondegenerate, then r≥2r \ge2 and

ω(β)≥p2(r−1)+1,\omega(\beta) \ge p^2(r - 1) + 1,

where ω(β)\omega(\beta) is the largest size of a set with pairwise nonzero β\beta-values.

Proof. Choose μ∈W∗\mu\in W^* independent of λ\lambda and put c=μ∘βc = \mu\circ\beta. Define TT by c(v,x)=b(Tv,x)c(v,x) = b(Tv,x). Since the values of β\beta span WW, TT is not scalar. Hence some v,Tvv,Tv are linearly independent: an operator preserving every line is scalar, as a basis and its pairwise sums show. On U=v⊥bU = v^{\perp_b} the functional x↦c(v,x)x \mapsto c(v,x) is therefore nonzero and descends to S=U/⟨v⟩S = U/\langle v\rangle. This space is symplectic of dimension 2r−22r - 2, so r≥2r \ge2.

By Lemma 3.1, SS has a scalar clique of size p(r−1)+1p(r-1)+1 meeting the kernel hyperplane once. Lift its vertices to x0,x1,…,xp(r−1)∈Ux_0,x_1,\ldots,x_{p(r-1)} \in U, with x0x_0 exceptional. Replace each xix_i, i>0i>0, by the affine line xi+Fpvx_i+\mathbb{F}_p v. Within a line, distinct vectors have nonzero β\beta-pairing because c(v,xi)≠0c(v,x_i) \ne0. Between different lines, and with x0x_0, the bb-pairings remain the nonzero quotient pairings. The resulting clique has 1+p⋅p(r−1)1+p\cdot p(r-1) vertices.

Scalar fibres and Sylow budgets

Scalarization applies a homomorphism with values in Fp\mathbb{F}_p to the commutator. A nonzero scalar value certifies noncommutation, but a zero value need not certify commutation. The radical consists of those elements invisible to the chosen scalar form; we must therefore count cliques inside its cosets as well as between them.

Let PP be a nonabelian pp-group. A nonzero homomorphism χ:P′/γ3(P)→Fp\chi:P'/\gamma_3(P)\to\mathbb{F}_p exists by nilpotence. We also write χ\chi for its composite with P′→P′/γ3(P)P'\to P'/\gamma_3(P). The pairing

bχ(x,y)=χ([x,y])b_\chi(x,y)=\chi([x,y])

is biadditive and conjugation-invariant. Its radical is

R=Rχ={x∈P:bχ(x,y)=0 for all y∈P}.R=R_\chi=\{x\in P:b_\chi(x,y)=0\text{ for all }y\in P\}.

It is normal and contains P′PpZ(P)P'P^pZ(P). The induced form on V=P/RV=P/R is nondegenerate, making VV an elementary abelian symplectic space. All scalarizations below use a nonzero χ\chi. We use the parameters

d=dim⁡FpV=2r,A=Z(R),m=log⁡p[R:A],s=log⁡p[A:Z(P)],n=log⁡pz(P)=d+m+s,h=log⁡p∣[A,P]∣.(8)\begin{aligned} d&=\dim_{\mathbb{F}_p}V=2r,\qquad A=Z(R),\qquad m=\log_p[R:A],\qquad s=\log_p[A:Z(P)],\\ n&=\log_p z(P)=d+m+s,\qquad h=\log_p|[A,P]|. \tag*{(8)} \end{aligned}

Since RR is normal in PP, its centre A=Z(R)A=Z(R) is normal and abelian. The identity n=d+m+sn=d+m+s separates the scalar quotient from two parts hidden in its radical: noncommutativity inside RR, measured by mm, and the excess of Z(R)Z(R) over Z(P)Z(P), measured by ss. The term hh records the action of PP on Z(R)Z(R). Here m=0m=0 or m≥2m\ge2, since a nontrivial central quotient is not cyclic, and [A,P][A,P] is the subgroup generated by the commutators. We call a coset gRgR a scalar fibre. Fibres above distinct vertices of a scalar clique are completely noncommuting with one another, so their internal clique sizes can be added.

The odd-prime estimates apply in every nilpotency class; compare the restricted families in [2, 14, 18].

Theorem 3.3 (Sylow budgets). Let PP be a finite pp-group, and put n=log⁡pz(P)n=\log_p z(P). For nonabelian PP, use the scalarization parameters in (3.1).

(i) If p=2p=2, every scalarization of nonabelian PP satisfies

ω(P)≥n+1+dm+h.(9)\omega(P)\ge n+1+dm+h. \tag*{(9)}

For nonabelian PP, equality ω(P)=n+1\omega(P)=n+1 forces PP to have class two and P/Z(P)P/Z(P) to be elementary abelian of even dimension. Every 22-group consequently satisfies

z(P)≤Lω(P).(10)z(P)\le L_{\omega(P)}. \tag*{(10)}

(ii) For odd pp,

ω(P)−1≥{pn/2,p≥5,2n−2,p=3.(11)\omega(P)-1 \ge\begin{cases} p^{n/2}, & p \ge5, \\ 2n-2, & p=3. \end{cases} \tag*{(11)}

In particular, z(P)≤2ω(P)−1z(P) \le2^{\omega(P)-1} for p≥5p \ge5, and z(P)≤κ2ω(P)−1z(P) \le\kappa2^{\omega(P)-1} for p=3p=3.

(iii) If p=3p=3 and z(P)=27z(P)=27, then ω(P)≥10\omega(P)\ge10. If p=3p=3, z(P)=81z(P)=81 and ω(P)≤8\omega(P)\le8, then P′≅C3P' \cong C_3, P/Z(P)≅F34P/Z(P)\cong\mathbb{F}_3^4 and ω(P)=7\omega(P)=7.

Proof. The abelian case is immediate, so assume PP nonabelian. The common fibre estimate below is the input for all three parts. The binary case uses a relation between the conjugation actions; the odd cases use lower bounds for the sizes of the fibres.

Common fibre estimate. Choose lifts gvg_v of a spanning scalar clique D⊂P/RD\subset P/R and put δv=log⁡p[A:CA(gv)]\delta_v=\log_p[A:C_A(g_v)]. The coset bounds give ω(gvR)≥m+δv+1\omega(g_vR)\ge m+\delta_v+1 and ω(gvR)≥ω(gvA)=pδv\omega(g_vR)\ge\omega(g_vA)=p^{\delta_v}. Distinct fibres above DD are completely adjacent. Since AA is normal and abelian, [a,gv]∈A[a,g_v]\in A and [aa′,gv]=[a,gv][a′,gv][aa',g_v]=[a,g_v][a',g_v] for a,a′∈Aa,a'\in A. Thus the maps a↦[a,gv]a\mapsto[a,g_v] are homomorphisms, with common kernel Z(P)Z(P), because the lifts and RR generate PP and RR centralizes AA. Their joint image has order psp^s and embeds in the product of their images, of orders pδvp^{\delta_v}. Hence

∑v∈Dδv≥s,ω(P)≥∑v∈Dmax⁡{m+δv+1,pδv}.(12)\sum_{v\in D}\delta_v\ge s,\qquad\omega(P)\ge\sum_{v\in D}\max\{m+\delta_v+1,p^{\delta_v}\}. \tag*{(12)}

Part (i): binary interaction. For p=2p=2, choose a zero-sum scalar clique of size d+1d+1 and lift it to g1,…,gd+1g_1,\ldots,g_{d+1}; put δi=log⁡2[A:CA(gi)]\delta_i=\log_2[A:C_A(g_i)]. Their actions αi\alpha_i on AA commute, since RR centralizes AA, and ∏iαi=1\prod_i\alpha_i=1 because ∏igi∈R\prod_i g_i\in R. Write AA additively and put Ai=(αi−1)AA_i=(\alpha_i-1)A and βi=α1⋯αi\beta_i=\alpha_1\cdots\alpha_i, with β0=1\beta_0=1. The maps

Φ:A⟶⨁iAi,a⟼((αi−1)a)i;T:⨁iAi⟶A,(vi)⟼∑iβi−1vi\Phi:A\longrightarrow\bigoplus_i A_i,\quad a\longmapsto((\alpha_i-1)a)_i;\qquad T:\bigoplus_i A_i\longrightarrow A,\quad(v_i)\longmapsto\sum_i\beta_{i-1}v_i

satisfy ker⁡Φ=Z(P)\ker\Phi=Z(P) and TΦ=0T\Phi=0 by telescoping. Each AiA_i is invariant under all βj\beta_j, so im⁡T=∑iAi=[A,P]\operatorname{im}T=\sum_i A_i=[A,P]. Counting im⁡Φ≤ker⁡T\operatorname{im}\Phi\le\ker T gives

∑iδi≥s+h.(13)\sum_i\delta_i\ge s+h. \tag*{(13)}

Substitution in (12) yields ω(P)≥(d+1)(m+1)+s+h=n+1+dm+h\omega(P)\ge(d+1)(m+1)+s+h=n+1+dm+h. If ω(P)=n+1\omega(P)=n+1, then m=h=0m=h=0, so R=A=Z(P)R=A=Z(P). Thus P/Z(P)P/Z(P) is elementary abelian of even dimension and P′≤Z(P)P'\le Z(P). In general n≤ω(P)−1n\le\omega(P)-1; when nn is odd, equality is impossible, so n≤ω(P)−2n\le\omega(P)-2. This proves (10).

Part (ii): odd-prime budgets. Let pp be odd and set μ=m+1\mu=m+1, so μ=1\mu=1 or μ≥3\mu\ge3. Choose DD maximum, hence spanning. Its size is at least pr+1p^r+1; for p=3p=3 it is at least 4r−14r-1, and is four when r=1r=1. For integers δ≥0\delta\ge0, put fp(μ,δ)=max⁡{μ+δ,pδ}f_p(\mu,\delta)=\max\{\mu+\delta,p^\delta\}. Then

fp(μ,δ)≥μ+12+p2δ,f3(μ,δ)≥μ+12+2δ(μ=1 or μ≥3).f_p(\mu,\delta)\ge\frac{\mu+1}{2}+\frac{p}{2}\delta,\qquad f_3(\mu,\delta)\ge\frac{\mu+1}{2}+2\delta\quad(\mu=1\text{ or }\mu\ge3).

The first inequality averages the two arguments of the maximum, using pδ≥1+(p−1)δp^\delta\ge1+(p-1)\delta. For the second, check δ=0,1\delta=0,1 directly; for δ≥2\delta\ge2, use the same averaging with 3δ≥1+3δ3^\delta\ge1+3\delta. Together with (12), these give, for p≥5p\ge5,

ω(P)≥1+pn2+(μ−1)(p(r−1)+1)2.\omega(P)\ge1+\frac{pn}{2}+\frac{(\mu-1)(p(r-1)+1)}{2}.

For p=3p=3, they give

ω(P)≥{2n,r=1,2n−1+(μ−1)(4r−5)2,r≥2.\omega(P)\ge\begin{cases} 2n, & r=1,\\ 2n-1+\dfrac{(\mu-1)(4r-5)}{2}, & r\ge2. \end{cases}

These prove (11). For p≥5p \ge5, p/2>log⁡2pp/2 > \log_{2} p gives z(P)≤2ω(P)−1z(P) \le2^{\omega(P)-1}. For p=3p = 3 and n≥4n \ge4,

z(P)2ω(P)−1≤4(34)n≤8164.\frac{z(P)}{2^{\omega(P)-1}} \le4\left(\frac{3}{4}\right)^n \le\frac{81}{64}.

When n=2n = 2, the scalar-plane clique gives ω(P)≥4\omega(P) \ge4, hence ratio at most 9/89/8. When n=3n = 3, necessarily r=1r = 1, so ω(P)≥6\omega(P) \ge6 and the ratio is at most 27/3227/32. The case n=1n = 1 is impossible because a nontrivial central quotient cannot be cyclic. This completes (ii).

Part (iii): the small ternary gap. Suppose p=3p = 3, [P:R]=9[P:R] = 9 and R≠Z(P)R \ne Z(P). Choose one nonzero representative of each of the four lines of P/RP/R. If RR is nonabelian, the nonabelian-coset bound gives four vertices in each of the four completely adjacent fibres, hence sixteen. If RR is abelian, the action of P/RP/R on RR is nontrivial. Its kernel is a proper subspace containing at most one of our four representatives. Each of the other three cosets has a clique of size at least three; the fourth has at least one. Thus ω(P)≥10\omega(P) \ge10 in either case.

At central index 2727 every nonzero scalarization has rank two and a noncentral radical, proving the first assertion of (iii). At index 8181 with ω(P)≤8\omega(P) \le8, the same argument excludes scalar radicals of index nine. Hence every scalar radical is the centre, and PP has class two with P/Z(P)≅F34P/Z(P) \cong\mathbb{F}_3^4. Also P′P' is elementary abelian, since [x,y]3=[x3,y]=1[x,y]^3 = [x^3,y] = 1. If its dimension were at least two, Lemma 3.2, with p=3p = 3 and r=2r = 2, would give ten vertices. Thus the commutator is a nondegenerate scalar form; its clique number is seven by the W(3,3)W(3,3) calculation recalled in § 1. This proves (iii). □

Combining the Sylow budgets

Proposition 3.4 (The full nilpotent profile). If FF is nilpotent and ω(F)≤k\omega(F) \le k, then z(F)≤M(k)z(F) \le M(k). Every value M(k)M(k) is attained by a nilpotent group. For k≥9k \ge9, the comparison with LkL_k is strict if an odd-order Sylow subgroup of FF is nonabelian.

Proof. The abelian case is immediate. Write F=∏pF(p)F = \prod_p F_{(p)} and put tp=ω(F(p))t_p = \omega(F_{(p)}). Cartesian products of cliques give ω(F)≥∏ptp\omega(F) \ge\prod_p t_p. The Sylow budgets give

z(F)≤κ2∑p(tp−1),z(F) \le\kappa2^{\sum_p(t_p-1)},

where the factor κ\kappa is needed only in the ternary part.

First let k=2a+1k = 2a + 1 or 2a+22a + 2, with a≥4a \ge4. If exactly one Sylow subgroup is nonabelian, put n=log⁡pz(F)n = \log_p z(F) for its prime pp. The binary bound gives z(F)≤Lkz(F) \le L_k when p=2p = 2. For p=3p = 3, the odd budget gives n≤a+1n \le a + 1, and 3a+1<4a3^{a+1} < 4^a. For p≥7p \ge7,

log⁡2z(F)≤(4a+2)log⁡2pp<12a+67≤2a.\log_2 z(F) \le\frac{(4a+2)\log_2 p}{p} < \frac{12a+6}{7} \le2a.

For p=5p = 5, 53<275^3 < 2^7 gives log⁡2z(F)<(28a+14)/15≤2a\log_2 z(F) < (28a+14)/15 \le2a for a≥7a \ge7; for a=4,5,6a = 4,5,6, the respective bounds z(F)≤125,625,3125<4az(F) \le125,625,3125 < 4^a finish the comparison. If at least two Sylow subgroups are nonabelian, their clique numbers are at least three and one is at least four. Hence

∏ptp−∑p(tp−1)≥7,z(F)≤κ2ω(F)−7<2k−2≤Lk.\prod_p t_p-\sum_p(t_p-1)\ge7,\qquad z(F)\le\kappa2^{\omega(F)-7}<2^{k-2}\le L_k.

The small quotients in Lemma 2.4 handle k≤4k \le4. For k=5,6,8k = 5,6,8, at most one Sylow subgroup is nonabelian, since two would already supply at least twelve vertices. The binary bound gives respectively 16,16,6416,16,64. For p=3p = 3, the odd budget leaves central exponents at most 3,3,43,3,4; exponent three is excluded by Theorem 3.3(iii). The resulting ternary maxima are 9,9,819,9,81. For a nonabelian Sylow subgroup P=F(p)P = F_{(p)} with p≥5p \ge5, the inequalities ω(P)−1≥plog⁡pz(P)/2\omega(P)-1 \ge p\log_p z(P)/2 and ω(P)≥p+1\omega(P) \ge p+1 leave 2525 at p=5p = 5, and 4949 at p=7p = 7 only for the cap eight. Thus the maxima at 5,6,85,6,8 are 16,25,8116,25,81. The cap seven follows from eight and the ternary example. The models of § 1 attain every bound. □

Binary rigidity

The numerical profile leaves one extra vertex at even binary caps. The next theorem rules it out unless the commutator form is scalar. Its proof uses only the bounds already established.

Theorem 3.5 (Uniform rigidity at a binary-square index). Let GG be finite and suppose z(G)=2nz(G)=2^n, where nn is even and n≥6n\ge6. Exactly one of the following holds:

G′≅C2,G/Z(G)≅F2n,ω(G)=n+1;ω(G)≥n+3.\begin{aligned} G' &\cong C_2,\quad G/Z(G)\cong\mathbb{F}_2^n,\quad\omega(G)=n+1;\\ \omega(G)&\ge n+3. \end{aligned}

In particular ω(G)=n+2\omega(G)=n+2 is impossible.

Proof. Every odd-order element lies in Z(G)Z(G), so G=P×OG=P\times O with PP a 2-group and OO central of odd order. Thus we may replace GG by PP. Suppose ω(P)≤n+2\omega(P)\le n+2. We first show that every scalar radical is Z(P)Z(P); affine inflation will then force the derived group to have order two. For any scalar radical RR, the interaction bound gives dm+h≤1dm+h\le1, hence m=0m=0 and h≤1h\le1.

Excluding a noncentral radical. If R>Z(P)R>Z(P), then RR is abelian and [R,P]=⟨c⟩[R,P]=\langle c\rangle has order two, hence is central. For u∈Ru\in R, the identity [u2,g]=[u,g]2=1[u^2,g]=[u,g]^2=1 gives R2≤Z(P)R^2\le Z(P). Thus the pairing

R/Z(P)×V⟶⟨c⟩,(uZ(P),gR)⟼[u,g],V=P/R,R/Z(P)\times V\longrightarrow\langle c\rangle,\qquad(uZ(P),gR)\longmapsto[u,g],\qquad V=P/R,

is bilinear with trivial left radical, so 1≤s≤d1\le s\le d. The right radical of this pairing is a proper subspace of VV. Choose a hyperplane containing it and a scalar clique of size d+1d+1 meeting that hyperplane once. Each of the other dd vertices acts nontrivially on RR, so its fibre contains at least two noncommuting elements. The remaining fibre contributes one, giving ω(P)≥2d+1\omega(P)\ge2d+1. If s≤d−2s\le d-2, this is at least n+3n+3. Since n=d+sn=d+s and both n,dn,d are even, the remaining case is s=ds=d, making the pairing perfect and n=2dn=2d, d≥4d\ge4.

Choose a symplectic clique a1,…,ad+1a_1,\ldots,a_{d+1} in VV and v∉{0,a1,…,ad+1}v\notin\{0,a_1,\ldots,a_{d+1}\}. Let v1,…,v2d+3v_1,\ldots,v_{2d+3} consist of two copies of each aja_j and one copy of vv, and lift them to gi∈Pg_i\in P. The perfect pairing realizes every fi∈V∗f_i\in V^* by bi∈Rb_i\in R with [bi,g]=cfi(gR)[b_i,g]=c^{f_i(gR)}. Moreover [R,P]≤Z(P)[R,P]\le Z(P) implies [R,P′]=1[R,P']=1: for u∈Ru\in R, the map g↦[u,g]g\mapsto[u,g] is a homomorphism to the centre and therefore kills P′P'. Hence

[gibi,gjbj]=[gi,gj]cfi(vj)+fj(vi).[g_i b_i,g_j b_j]=[g_i,g_j]c^{f_i(v_j)+f_j(v_i)}.

Nonzero scalar pairings remain nontrivial because c∈[R,P]c\in[R,P] and RR is the scalar radical, so χ(c)=0\chi(c)=0. For every zero scalar pairing whose commutator lies in ⟨c⟩\langle c\rangle, prescribe its correction exponent to make the displayed commutator equal to cc; prescribe zero on the other zero pairings. To meet the prescriptions, denote the functionals for the two copies of aja_j by fj,1,fj,2f_{j,1},f_{j,2}, and set the functional for vv equal to zero. Distinct doubled families have scalar pairing one, so there is no constraint between them. Within the jjth family the only constraint is

fj,1(aj)+fj,2(aj)=εj,f_{j,1}(a_j)+f_{j,2}(a_j)=\varepsilon_j,

where εj∈F2\varepsilon_j\in\mathbb{F}_2 is the prescribed correction for its repeated pair. If bχ(aj,v)=0b_\chi(a_j,v)=0, the two additional constraints are fj,ℓ(v)=ηj,ℓf_{j,\ell}(v)=\eta_{j,\ell} for ℓ=1,2\ell=1,2, where ηj,ℓ∈F2\eta_{j,\ell}\in\mathbb{F}_2 are the prescribed corrections for the pairs with vv. If bχ(aj,v)≠0b_\chi(a_j,v)\ne0, there is no constraint on these values. Choose fj,1(aj)=0f_{j,1}(a_j)=0 and fj,2(aj)=εjf_{j,2}(a_j)=\varepsilon_j. Since aja_j and vv are linearly independent, each choice of the required values on these two vectors extends to a functional on VV. The choices for distinct families are independent, so all prescriptions hold simultaneously. The corrected lifts therefore form a clique of size 2d+3=n+32d+3=n+3, again a contradiction.

Identifying the commutator form. Thus every scalar radical is Z(P)Z(P). Since it contains P′P2P'P^2, the group has class two and elementary abelian central quotient; P′P' is also elementary abelian. If dim⁡F2P′≥2\dim_{\mathbb{F}_2}P'\ge2, Lemma 3.2 gives ω(P)≥2n−3≥n+3\omega(P)\ge2n-3\ge n+3. Consequently P′≅C2P'\cong C_2, and its nondegenerate symplectic form gives ω(P)=n+1\omega(P)=n+1. The two alternatives are exclusive, as claimed. □

Remark 3.6 (The dimension threshold). The restriction n≥6n \ge6 is necessary: H=UT3(F4)H = \mathrm{UT}_3(\mathbb{F}_4) has H′=Z(H)≅C22H' = Z(H) \cong C_2^2, z(H)=16z(H) = 16 and ω(H)=5\omega(H) = 5. Indeed, its commutator is the determinant of the two coordinates in F42\mathbb{F}_4^2, so a maximum clique has one representative from each of the five points of P1(F4)\mathbb{P}^1(\mathbb{F}_4). Thus the scalar dichotomy fails when n=4n = 4.

Centralizer reduction and saturation

The nilpotent case is settled. To control a general group, we must bound a centralizer index and the centrality lost outside that subgroup. We first quantify the loss, then select the centralizer and obtain the numerical bound used later. Two saturation principles will exclude the finite cases left by that bound.

Charging a loss of centrality

For H≤GH \le G and Z(G)≤A≤Z(H)Z(G) \le A \le Z(H), the index [A:Z(G)][A : Z(G)] measures how much of AA is central in HH but not in GG. Adjoining generators from HH up to GG removes this excess. At each nontrivial step, Lemma 2.3 forces new clique vertices. No normality of AA in GG is required.

For a positive integer bb, let vp(b)v_p(b) be the exponent of pp in bb. Define the additive prime cost and, for integers D≥0D \ge0, its product envelope by

σ(b)=∑pp vp(b),σ(1)=0,P(D)=max⁡{b∈Z≥1:σ(b)≤D}.\sigma(b) = \sum_p p\,v_p(b), \qquad\sigma(1) = 0, \qquad\mathcal{P}(D) = \max\{b \in\mathbb{Z}_{\ge1} : \sigma(b) \le D\}.

Since σ(uv)=σ(u)+σ(v)\sigma(uv) = \sigma(u) + \sigma(v), the envelope satisfies P(D)P(E)≤P(D+E)\mathcal{P}(D)\mathcal{P}(E) \le\mathcal{P}(D + E) for D,E≥0D,E \ge0. Splitting a composite factor never increases its additive cost, since a+b≤aba + b \le ab for a,b≥2a,b \ge2. Thus σ(b)\sigma(b) is the minimum sum of nontrivial factors with product bb. Integer break gives P(0)=P(1)=1\mathcal{P}(0) = \mathcal{P}(1) = 1 and, for D≥2D \ge2, the usual product of 2’s and 3’s: 3D/33^{D/3} if D≡0D \equiv0, 4⋅3(D−4)/34 \cdot3^{(D-4)/3} if D≡1D \equiv1, and 2⋅3(D−2)/32 \cdot3^{(D-2)/3} if D≡2(mod3)D \equiv2 \pmod3.

Theorem 4.1 (Central-loss calculus). Let H≤GH \le G, Z≤A≤Z(H)Z \le A \le Z(H) and put b=[A:Z]b = [A : Z]. Then

ω(G)−ω(H)≥σ(b).(14)\omega(G) - \omega(H) \ge\sigma(b). \tag*{(14)}

More explicitly, for G=⟨H,g1,…,gt⟩G = \langle H,g_1,\ldots,g_t\rangle, put H0=HH_0 = H, Hi=⟨H,g1,…,gi⟩H_i = \langle H,g_1,\ldots,g_i\rangle, Ai=A∩Z(Hi)A_i = A \cap Z(H_i) for 0≤i≤t0 \le i \le t, and di=[Ai−1:Ai]d_i = [A_{i-1} : A_i] for 1≤i≤t1 \le i \le t. Then

∑di>1di≤N−ω(H),∏idi=b.(15)\sum_{d_i>1} d_i \le N - \omega(H), \qquad\prod_i d_i = b. \tag*{(15)}

For distinct primes p,q∣bp,q \mid b, one moreover has

N−ω(H)≥σ(b)+pq−p−q.(16)N - \omega(H) \ge\sigma(b) + pq - p - q. \tag*{(16)}

If b=36b = 36, this lower bound improves to twelve. No normality of AA in GG is assumed.

Proof. Since Ai=CAi−1(gi)A_i = C_{A_{i-1}}(g_i), twisting adds at least did_i vertices at each nontrivial step. Sum these increases and use At=ZA_t = Z. Splitting the factors did_i into primes proves (14).

For the surcharge choose gg outside the union of the two proper subgroups CG(A(p))C_G(A_{(p)}) and CG(A(q))C_G(A_{(q)}). Then d=[A:CA(g)]d = [A : C_A(g)] is divisible by pqpq. Twisting first by gg and applying the ordinary cost to the remaining loss gives N−ω(H)≥d+σ(b/d)N - \omega(H) \ge d + \sigma(b/d). Splitting dd while retaining pqpq as one factor gives d≥pq+σ(d/pq)d \ge pq + \sigma(d/pq), proving (16). When b=36b = 36, the first mixed index is 6, 12, 18 or 36. Only six could cost less than twelve; it leaves a loss six, whose mixed surcharge costs six more. ∎

Every central loss with available budget DD satisfies

b≤P(D)≤3D/3.(17)b \le\mathcal{P}(D) \le3^{D/3}. \tag*{(17)}

We set P(D)=0\mathcal{P}(D)=0 for D<0D<0. At budgets D=5,7,11D=5,7,11, the mixed surcharge sharpens this to b≤5,9,36b\le5,9,36, respectively: mixed losses have σ(b)≤D−1\sigma(b)\le D-1, while the largest pure prime powers of cost at most DD are 5,9,325,9,32.

For later use, put K=CG(L)K=C_G(L), a=[G:K]a=[G:K], b=[Z(K):Z]b=[Z(K):Z] and q=ω(L/Z(L))q=\omega(L/Z(L)). Then

z(G)=a z(K)b,N≥qω(K)+σ(b).(18)z(G)=a\,z(K)b,\qquad N\ge q\omega(K)+\sigma(b). \tag*{(18)}

Indeed Z(KL)=Z(K)Z(KL)=Z(K): a central element of KLKL centralizes LL and hence belongs to KK. Apply (7) and the central-loss bound with H=KLH=KL, A=Z(K)A=Z(K).

Fourfold centralizer selection

Lemma 4.2 (Complement rigidity). If k≥1k\ge1 is an integer, D<GD<G and ω(G∖D)≤k\omega(G\setminus D)\le k, then either CG(x)C_G(x) is abelian of index at most kk for some x∉Dx\notin D, or some x∉Dx\notin D satisfies [G:CG(x)]≤k−1[G:C_G(x)]\le k-1.

Proof. Suppose all those indices are at least kk. Put S=G∖DS=G\setminus D and d=[G:D]d=[G:D]. For C=CG(x)C=C_G(x), x∈Sx\in S, one has q=[C:C∩D]≤dq=[C:C\cap D]\le d and hence

∣C∩S∣=∣C∣(1−1/q)≤∣S∣/k.|C\cap S|=|C|(1-1/q)\le|S|/k.

A maximum clique in SS covers SS by at most kk such sets. Thus exactly kk sets partition SS, each with centralizer of index kk. The elements of each part commute pairwise: otherwise two of its vertices could replace its selected vertex, enlarging the clique. Moreover C∖DC\setminus D generates CC, since for s∈C∖Ds\in C\setminus D and a∈C∩Da\in C\cap D, both s,sas,sa belong to that part. Thus CC is abelian.

Proposition 4.3 (Fourfold descent). If GG is nonnilpotent, one of the following holds:

(A) There is x∉F(G)x\notin F(G) such that CG(x)C_G(x) is abelian and [G:CG(x)]≤N[G:C_G(x)]\le N.

(B) There are conjugate elements x,yx,y, with x∉F(G)x\notin F(G), such that L=⟨x,y⟩L=\langle x,y\rangle is nonnilpotent. For K=CG(L)K=C_G(L), r=∣xG∣r=|x^G| and a=[G:K]a=[G:K],

r≤N−1,r∣a≤r(r−1),ω(KL)≥4ω(K).(19)r\le N-1,\qquad r\mid a\le r(r-1),\qquad\omega(KL)\ge4\omega(K). \tag*{(19)}

Here F(G)F(G) denotes the Fitting subgroup.

Proof. Apply Lemma 4.2 to D=F(G)D=F(G). Outside its abelian alternative, choose x∉F(G)x\notin F(G) with r≤N−1r\le N-1. Baer–Suzuki in its nilpotent formulation [16 p. 281] supplies a conjugate yy for which L=⟨x,y⟩L=\langle x,y\rangle is nonnilpotent. The group CG(x)C_G(x) fixes xx in xGx^G, so its orbit on yy has size at most r−1r-1. Its stabilizer is KK, proving the index bounds. Finally L/Z(L)L/Z(L) is nonnilpotent and has a four-point clique by Lemma 2.4. Apply (7).

Losses for an abelian centralizer

Lemma 4.4 (Abelian centralizer losses). Let LL be nonnilpotent, K=CG(L)K=C_G(L) abelian, and put C=CG(K)C=C_G(K), a=[G:K]a=[G:K], b=[K:Z]b=[K:Z] and h=ω(C)h=\omega(C). Then Z(C)=KZ(C)=K, C/KC/K is nonnilpotent, and

h≥4,h+σ(b)≤N.(20)h\ge4,\qquad h+\sigma(b)\le N. \tag*{(20)}

If h=4h=4, then C/K≅S3C/K\cong S_3, 6∣a6\mid a, and CC has an abelian subgroup of loss 3b3b. Consequently, for N≤kN\le k and k≥5k\ge5,

b≤J(k):=max⁡{P(k−5),3⌊(k−4)/3⌋};N≤12⇒b≤9.(21)b\le J(k):=\max\{\mathcal{P}(k-5),3^{\lfloor(k-4)/3\rfloor}\};\qquad N\le12\Rightarrow b\le9. \tag*{(21)}

Proof. Since L≤CL \le C, a central element of CC lies in CG(L)=KC_G(L)=K; the reverse inclusion follows from the definition of CC. Also C/KC/K contains the nonnilpotent subgroup L/Z(L)L/Z(L), so h≥4h \ge4. Apply the central-loss budget to K≤Z(C)K \le Z(C). If h=4h=4, Lemma 2.4 leaves only S3S_3. The preimage of its subgroup of order three is abelian, since its quotient by the central subgroup KK is cyclic.

For h≥5h \ge5, b≤P(k−5)b \le P(k-5). For h=4h=4, the abelian loss 3b3b has budget k−1k-1. If bb is a power of three, its ordinary cost gives the second term in (21). Otherwise the mixed surcharge gives σ(b)+4≤k−1\sigma(b)+4 \le k-1, and hence the first term. At cap twelve the budget-seven refinement gives b≤9b \le9 when h≥5h \ge5. For h=4h=4, only b=10,11,12b=10,11,12 remain to exclude: the losses 30,33,3630,33,36 cost at least 17,33,1217,33,12, exceeding eleven.

Near-minimal ternary losses

The following theorem uses the prime-index result of Bergman and Lenstra. A noncentral commutator subgroup of order three would produce a local twelve-point clique; the central cost of the remaining dimensions then exceeds the available budget. This is the divisibility input for the ternary losses left by Lemma 4.4.

Theorem 4.5 (Near-minimal ternary losses). Let Z≤A≤GZ \le A \le G be abelian, with [A:Z]=2ε3r[A:Z]=2^{\varepsilon}3^r, r≥2r \ge2 and ε∈{0,1}\varepsilon\in\{0,1\}. If N≤3r+3+2εN \le3r+3+2\varepsilon, then 3r∣[G:A]3^r \mid[G:A].

Proof. First force orbits of size at most three. Put D=ZA(3)D=ZA_{(3)}. If [D:CD(g)]=3e≥9[D:C_D(g)]=3^e \ge9, taking gg first in the central-loss sequence for AA and splitting off any factor two costs at least

3e+3(r−e)+2ε≥3r+3+2ε≥N,3^e+3(r-e)+2\varepsilon\ge3r+3+2\varepsilon\ge N,

exceeding its budget N−1N-1; the splitting uses 2⋅3e≥3e+22 \cdot3^e \ge3^e+2 if the binary part also moves. Thus every DD-orbit in GG has size one or three.

Apply Bergman–Lenstra. Put E=D/ZE=D/Z and T=G⋊ET=G \rtimes E, where dZ∈EdZ \in E acts on GG by the inner automorphism induced by dd. This action is well defined and faithful, since its kernel is CD(G)/Z=1C_D(G)/Z=1. We identify GG and EE with the normal factor and the complement in TT, respectively. If a normal subgroup VV of TT is contained in EE, then [V,G]≤V∩G=1[V,G] \le V \cap G=1; faithfulness gives V=1V=1. Thus EE has trivial core in TT. For g∈Gg \in G, the orbit bound above gives [E:E∩Eg]=[D:CD(g)]∈{1,3}[E:E \cap E^g]=[D:C_D(g)] \in\{1,3\}. Since EE normalizes itself and T=GET=GE, these are also all the possible indices for conjugation by TT. Both values occur: otherwise the action of the nontrivial group EE on GG would be trivial. By [3 Theorem 2], a normal subgroup MM of TT either lies in EE with E/ME/M a transitive permutation group of degree three, or contains EE with [M:E]=3[M:E]=3. The first case forces M=1M=1 and ∣E∣=3|E|=3, since a faithful abelian transitive group is regular, contrary to r≥2r \ge2. Thus the second holds, and U=M∩GU=M \cap G is normal of order three, since M=UEM=UE and [M:E]=3[M:E]=3. As E≤ME \le M and both MM and GG are normal in TT, we have [G,E]≤M∩G=U[G,E] \le M \cap G=U. This commutator subgroup is nontrivial by faithfulness, so it is all of UU. The action defining the semidirect product identifies it with [D,G][D,G] inside GG. Finally EE centralizes UU, because a 33-group acts trivially on C3C_3, whose automorphism group has order two. The same holds for conjugation by DD.

A local twelve-point clique forces centrality. We claim that U≤ZU \le Z. Otherwise CG(U)C_G(U) has index two; choose g∉CG(U)g \notin C_G(U). Since DD is abelian and centralizes UU, the group B=DUB=DU is abelian. Moreover [D,G]=U[D,G]=U and normality of UU give [B,G]=U≤B[B,G]=U \le B, so BB is normal. Also [B:Z]=3t[B:Z]=3^t for t∈{r,r+1}t \in\{r,r+1\}. There is u∈CG(U)u \in C_G(U) not centralizing BB: otherwise CB(g)=CB(G)=ZC_B(g)=C_B(G)=Z, whereas [B:CB(g)]=3[B:C_B(g)]=3 and t≥2t \ge2. The index-three subgroups CB(u),CB(g)C_B(u),C_B(g) intersect in B0B_0 with [B:B0]=9[B:B_0]=9, because only CB(u)C_B(u) contains UU. Put H0=⟨B,u⟩H_0=\langle B,u\rangle, H=⟨H0,g⟩H=\langle H_0,g\rangle and C=CH0(B)C=C_{H_0}(B). The action of H0H_0 on BB is generated by the order-three action of uu, so [H0:C]=3[H_0:C]=3; also B0≤Z(H)B_0 \le Z(H). For w∈gH0w \in gH_0, we have CH0(w)≤CC_{H_0}(w) \le C. Indeed, if h∈H0h \in H_0 commutes with ww, then for every b∈Bb \in B,

[b,h]=[bw,h]=[b,h]w=[b,h]−1.[b,h]=[b^w,h]=[b,h]^w=[b,h]^{-1}.

the first equality uses bw∈bUb^w \in bU and [U,H0]=1[U,H_0]=1, and ww inverts UU. Thus [b,h]=1[b,h]=1. Also [B:CB(w)]=3[B:C_B(w)]=3, so

[H0:CH0(w)]≥[H0:C][B:CB(w)]=9.[H_0:C_{H_0}(w)] \ge[H_0:C][B:C_B(w)]=9.

Since this index bound holds for every w∈gH0w \in gH_0, the averaging assertion in Lemma 2.2 gives a nine-point clique in gH0gH_0. Choose a∈Ba \in B with [u,a]≠1[u,a] \ne1. The three elements uajua^j, 0≤j<30 \le j < 3, form a clique in H0∖CH_0 \setminus C: the nontrivial commutator [u,a]∈U[u,a] \in U is central in H0H_0 and has order three. They are completely adjacent to gH0gH_0, because CH0(w)≤CC_{H_0}(w) \le C for every w∈gH0w \in gH_0. Hence ω(H)≥12\omega(H) \ge12. Applying the central-loss bound to B0≤Z(H)B_0 \le Z(H) now gives

N≥12+σ([B0:Z])=12+3(t−2)≥3r+6>N,N \ge12+\sigma([B_0:Z])=12+3(t-2)\ge3r+6>N,

a contradiction. Therefore U≤ZU \le Z, so DD is normal and [D3,G]=1[D^3,G]=1 makes D/Z≅F3D/Z \cong\mathbb{F}_3.

Read off the divisibility. The commutator map is the homomorphism

G⟶Hom⁡(D/Z,U),g⟼(dZ⟼[d,g]).G \longrightarrow\operatorname{Hom}(D/Z,U), \qquad g \longmapsto(dZ \longmapsto[d,g]).

Identify UU with the additive group of F3\mathbb{F}_3. The image is then a subspace of the dual of D/Z≅F3rD/Z \cong\mathbb{F}_3^r, with zero annihilator because CD(G)=ZC_D(G)=Z. A proper subspace of this dual has a nonzero annihilator, by the dimension formula for the evaluation pairing. Thus the image is the whole dual. Its kernel is CG(D)C_G(D), so [G:CG(D)]=3r[G:C_G(D)]=3^r. Since D≤CG(D)D \le C_G(D), we obtain 3r∣[G:D]=2ε[G:A]3^r \mid[G:D]=2^\varepsilon[G:A], proving the assertion.

Lifting the ternary loss. Return to the notation of Lemma 4.4, and put m=[C:K]m=[C:K]. If b=2ε3sb=2^\varepsilon3^s, ε∈{0,1}\varepsilon\in\{0,1\}, s≥1s \ge1, 3∣m3 \mid m and N≤3s+6+2εN \le3s+6+2\varepsilon, then

3s+2∣a.(22)3^{s+2} \mid a. \tag*{(22)}

For (22), lift an element of order three in C/KC/K to yy. Then A=K⟨y⟩A=K\langle y\rangle is abelian with loss 3b3b and index a/3a/3 in GG. Theorem 4.5 gives 3s+1∣a/33^{s+1}\mid a/3, as required.

Saturation of an affine conjugacy class

In this subsection we write the conjugation action on a class using left conjugation. Choose a bijection u↦xuu \mapsto x_u from a finite abelian group AA onto the class XX. The law is affine if

xuxvxu−1=x(1−f)u+fv,f∈Aut⁡(A),1−f invertible.x_u x_v x_u^{-1}=x_{(1-f)u+fv}, \qquad f \in\operatorname{Aut}(A),\quad1-f\text{ invertible}.

Thus u,vu,v are labels in AA, whereas xu,xvx_u,x_v are elements of GG. Such a class is connected: the subgroup it generates acts transitively on it.

Lemma 4.6 (Affine saturation). Suppose a conjugacy class XX of GG has the affine law above, ∣A∣=q>1|A|=q>1, and put R=⟨X⟩R=\langle X\rangle. Then ω(G)≥q+1\omega(G)\ge q+1. If ω(G)≤q+1\omega(G)\le q+1, then

CG(R)=Z(G),z(G)∣q∣Aut⁡(A)∣.C_G(R)=Z(G), \qquad z(G)\mid q|\operatorname{Aut}(A)|.

Proof. Conjugation identifies R/Z(R)R/Z(R) with the inner group of XX. For the connected affine law, the standard descriptions give R/Z(R)≅A⋊⟨f⟩R/Z(R)\cong A\rtimes\langle f\rangle and Aut⁡(X)=A⋊CAut⁡(A)(f)\operatorname{Aut}(X)=A\rtimes C_{\operatorname{Aut}(A)}(f); see [1 Lemma 1.9, Corollary 1.25]. The qq elements (a,f)(a,f), a∈Aa\in A, are pairwise noncommuting, and every (w,1)(w,1) with w≠0w\ne0 fails to commute with all of them, since 1−f1-f is invertible. Thus ω(R/Z(R))≥q+1\omega(R/Z(R))\ge q+1. Put E=CG(R)E=C_G(R). If N≤q+1N \le q+1, (18) gives

q+1≥N≥(q+1)ω(E)+σ([Z(E):Z]).q+1 \ge N \ge(q+1)\omega(E)+\sigma([Z(E):Z]).

Hence E=ZE=Z. Conjugation embeds G/ZG/Z in Aut⁡(X)\operatorname{Aut}(X), so z(G)∣q∣CAut⁡(A)(f)∣z(G)\mid q|C_{\operatorname{Aut}(A)}(f)|, which implies the required divisibility.

Corollary 4.7 (Small conjugacy classes). Let x∉F(G)x\notin F(G) and r=∣xG∣r=|x^G|. If r=pr=p is prime, then N≥p+1N\ge p+1; if also N≤p+1N\le p+1, then z(G)∣p(p−1)z(G)\mid p(p-1). If r=9r=9 and N≤10N\le10, either there is a conjugate yy with ⟨x,y⟩\langle x,y\rangle nonnilpotent and [G:CG(x,y)]≤18[G:C_G(x,y)]\le18, or z(G)∣54z(G)\mid54 or z(G)∣432z(G)\mid432.

Proof. Set X=xGX=x^G and R=⟨X⟩R=\langle X\rangle. Normality of RR makes all its orbits on XX equally large, so ∣xR∣∣r|x^R|\mid r. This size is not one: otherwise X⊆Z(R)X\subseteq Z(R), so RR is abelian and x∈F(G)x\in F(G). If r=9r=9 and ∣xR∣=3|x^R|=3, then F(R)≤F(G)F(R)\le F(G) by characteristicity. Baer–Suzuki in RR supplies y∈xRy\in x^R with ⟨x,y⟩\langle x,y\rangle nonnilpotent. Since CG(x)C_G(x) preserves xRx^R and fixes xx, [G:CG(x,y)]=9∣yCG(x)∣≤18[G:C_G(x,y)]=9|y^{C_G(x)}|\le18.

In the remaining cases, X=xRX=x^R is connected. The prime-order theorem of Etingof–Soloviev–Guralnick [13] and Graña’s prime-square theorem [15 Proposition 3.10] make its law affine. Apply Lemma 4.6. For prime pp, A=CpA=C_p has ∣Aut⁡(A)∣=p−1|\operatorname{Aut}(A)|=p-1; for nine, A=C9A=C_9 or C32C_3^2 has automorphism group of order six or 48, giving the two stated divisibilities.

Completion of the exact profile

We first establish the small caps and isolate the finite exceptions. Those exceptions are grouped by the mechanism that excludes them: ternary divisibility at caps eight, fourteen and sixteen, and affine saturation at cap ten. After this finite block, one numerical envelope completes the proof by induction.

The global upper bound

Proof of Theorem 1.1. The models in § 1 give every lower bound. Proposition 3.4 gives every nilpotent upper bound, and Lemma 2.4 settles all caps k≤4k\le4. We therefore consider nonnilpotent groups. It is enough to prove the bounds at cap five and at every even cap k≥6k\ge6: each odd cap 2j−1≥72j-1\ge7 then follows from cap 2j2j, since M(2j−1)=M(2j)M(2j-1)=M(2j).

Step 1. A common budget for the two cases. In case (A) of Proposition 4.3, put A=CG(x)A=C_G(x), a=[G:A]a=[G:A] and b=[A:Z]b=[A:Z]. Then

a≤k,σ(b)≤k−1,z(G)=ab.a\le k,\qquad\sigma(b)\le k-1,\qquad z(G)=ab.

In case (B), retain the selected class size rr and put a=[G:K]a=[G:K], t=ω(K)t=\omega(K) and b=[Z(K):Z]b=[Z(K):Z]. The accounting identity and fourfold amplification give

z(G)=a z(K)b,4t+σ(b)≤k,r∣a≤r(r−1),r≤k−1.(23)z(G)=a\,z(K)b,\qquad4t+\sigma(b)\le k,\qquad r\mid a\le r(r-1),\qquad r\le k-1. \tag*{(23)}

If KK is abelian, Lemma 4.4 gives b≤J(k)b\le J(k). If KK is nonabelian, then t≥3t\ge3. Whenever the profile is already known at tt, it gives z(K)≤M(t)z(K)\le M(t).

Here is the numerical fact that lets us absorb this internal index. For t=3,…,8t=3,\ldots,8, the respective values of σ(M(t))\sigma(M(t)) are 4,6,8,10,12,124,6,8,10,12,12; for t≥9t\ge9, the binary formula gives σ(M(t))=4⌊(t−1)/2⌋\sigma(M(t))=4\lfloor(t-1)/2\rfloor. Thus σ(M(t))≤2(t−1)\sigma(M(t))\le2(t-1) for every t≥3t\ge3, and

z(K)b≤M(t)P(k−4t)≤P(k−2t−2)≤P(k−8)≤J(k).(24)\begin{aligned} z(K)b\le M(t)\mathcal{P}(k-4t) \\ &\le\mathcal{P}(k-2t-2)\le\mathcal{P}(k-8)\le J(k). \tag*{(24)} \end{aligned}

The second line uses the additive prime cost of the numerical majorant in the first line. Consequently, once the bound at tt is available, every selected centralizer satisfies z(G)≤aJ(k)z(G)\le aJ(k). For all caps k≤16k \le16, we have t≤4t \le4, so the required bound at tt is already supplied by Lemma 2.4. For the later caps it will be supplied by induction.

Step 2. The base caps five and six. At cap five, the cases N≤4N \le4 are already settled. If N=5N = 5, Lemma 2.2 gives an irredundant cover by five centralizers with intersection ZZ, so z(G)≤16z(G) \le16 by [8 Theorem 1.2].

At cap six, case (B) has 4t≤64t \le6, so KK is abelian. Since J(6)=1J(6) = 1, this gives z(G)≤20<25z(G) \le20 < 25. In case (A), a≤6a \le6 and b≤5b \le5, so an excess over 2525 requires (a,b)=(6,5)(a,b) = (6,5).

For (a,b)=(6,5)(a,b) = (6,5), the quotient Q=G/ZQ = G/Z has order thirty and a normal Sylow 5-subgroup. Indeed, otherwise its six Sylow 5-subgroups supply twenty-four nonidentity elements, forcing the Sylow 3-subgroup to be normal. The quotient of order ten then gives a normal cyclic subgroup of order fifteen; its Sylow 5-subgroup is characteristic and therefore normal in QQ, a contradiction. Hence A/Z≅C5A/Z \cong C_5 is normal. Let HH be the kernel of the action of GG on A/ZA/Z. Since A=CG(x)A = C_G(x), the map

H/A⟶Z,hA⟼[x,h]H/A \longrightarrow Z,\qquad hA \longmapsto[x,h]

is injective, and its image has exponent five because x5∈Zx^5 \in Z. Its order also divides six, so H=AH = A. This embeds the group G/AG/A, of order six, in Aut⁡(C5)\operatorname{Aut}(C_5), of order four. Thus ζ(5)=M(5)=16\zeta(5) = M(5) = 16 and ζ(6)=M(6)=25\zeta(6) = M(6) = 25.

Step 3. The finite caps eight through sixteen. The following table records upper bounds for z(G)z(G) in each alternative. An entry at most M(k)M(k) closes that branch; a larger entry indicates a case still to be examined. A dash means that 4ω(K)≤k4\omega(K) \le k excludes a nonabelian KK. For case (A), use kP(k−1)kP(k - 1), sharpened at budgets seven and eleven to losses nine and thirty-six. For nonabelian KK, use (k−1)(k−2)P(k−8)(k - 1)(k - 2)P(k - 8) from (24). For abelian KK, use (k−1)(k−2)J(k)(k - 1)(k - 2)J(k), with the sharper b≤9b \le9 at cap twelve from Lemma 4.4.

kkM(k)M(k)case (A)case (B) KK nonabeliancase (B) KK abelian
88172—126
10256270—648
121024432440990
144096151214044212
16163843888378017010

Table 1. Preliminary bounds for the finite caps. The three right-hand columns distinguish case (A) and the two possibilities for KK in case (B); the entries are bounds, not attained values.

The case k=12k = 12 is therefore settled. In case (A) the only possible excess is at cap ten, where a≤10a \le10, b≤27b \le27 and ab>256ab > 256 require a=10a = 10 and b=26b = 26 or 2727. The loss 26 costs fifteen, while loss 27 is excluded by Theorem 4.5, which would give 27∣1027 \mid10. All remaining excesses thus have abelian selected KK.

We now treat caps eight, fourteen and sixteen together, and then return to ten. This order has no inductive dependency: throughout this finite block the only profile bounds used are those at caps at most five, already proved above.

Step 4. Three cases excluded by ternary divisibility. Assume z(G)>M(k)z(G) > M(k) in case (B), with KK abelian. At cap eight, if r=7r = 7, Corollary 4.7 gives z(G)∣42z(G) \mid42. Otherwise r≤6r \le6. Since b≤3b \le3 and r∣a≤r(r−1)r \mid a \le r(r - 1), the inequality ab>81ab > 81 then forces (a,b)=(30,3)(a,b) = (30,3).

At cap fourteen, a≤156a \le156, b≤27b \le27 and ab>4096ab > 4096 force b=27b = 27 and a≥152a \ge152. Hence r=13r = 13 and a=156a = 156. At cap sixteen, a≤210a \le210, b≤81b \le81 and ab>16384ab > 16384 leave 79≤b≤8179 \le b \le81. But σ(b)≤12\sigma(b) \le12 excludes 79 and 80, so b=81b = 81; the index constraints then give r=15r = 15 and a=210a = 210.

Put C=CG(K)C = C_G(K) and m=[C:K]m = [C : K]. At caps eight and fourteen, Lemma 4.4 gives ω(C)≤k−σ(b)=5\omega(C) \le k - \sigma(b) = 5. Since Z(C)=KZ(C) = K, the cap-five bound from Step 2 yields m≤16m \le16. Also mm and C/KC/K is nonnilpotent. For a=30a = 30 this leaves m=6m = 6 or 10; for a=156a = 156 it leaves m=6m = 6 or 12. The order-ten possibility is dihedral and has a six-point clique (five reflections and a nontrivial rotation), contradicting ω(C)≤5\omega(C) \le5. At cap sixteen, 4≤ω(C)≤16−12=44 \le\omega(C) \le16 - 12 = 4, so C/K≅S3C/K \cong S_3 and m=6m = 6.

In every remaining row, 3∣m3 \mid m. The ternary lifting criterion (22) now gives the contradictions in Table 2. The criterion requires respectively 8≤98 \le9, 14≤1514 \le15 and 16≤1816 \le18.

kkaabbmmrequired divisibility
8303627∣3027 \mid30
14156276 or 12243∣156243 \mid156
16210816729∣210729 \mid210

Table 2. Divisibility obstructions in the three ternary cases. Each divisibility is forced by (22) and is incompatible with the displayed value of aa.

Step 5. The remaining cases at cap ten. It remains to consider cap ten with KK abelian. Here a≤72a \le72, b≤9b \le9 and ab>256ab > 256, so b≥4b \ge4. Again put C=CG(K)C = C_G(K). Loss five is impossible: lift an element of prime order qq in C/KC/K. Together with KK, its lift generates an abelian subgroup of loss 5q5q. The central-loss calculus assigns this loss a cost of at least ten, exceeding the available budget nine.

For b=6,7,8,9b = 6, 7, 8, 9, the central-loss cost is at least six; for b=6b = 6 this uses the mixed surcharge, and for the other values the ordinary cost suffices. Thus ω(C)=4\omega(C) = 4, unless loss seven already contradicts 4+σ(b)≤104 + \sigma(b) \le10. The small-quotient lemma gives C/K≅S3C/K \cong S_3 and 6∣a6 \mid a. Lifting its subgroup of order three gives an abelian subgroup of loss 3b3b. If b=8b = 8, this is loss 24, whose mixed cost is ten, again impossible. If b=9b = 9, (22) gives 81∣a81 \mid a, contrary to a≤72a \le72. If b=6b = 6, the same criterion gives 27∣a27 \mid a; together with 6∣a6 \mid a, this forces a=54a = 54, and the class-index constraints force r=9r = 9.

Finally, if b=4b = 4, excess requires a≥65a \ge65. The constraints r≤9r \le9, r∣a≤r(r−1)r \mid a \le r(r - 1) then give r=9r = 9 and a=72a = 72. We have reduced the entire cap-ten problem to

(r,a,b,z(G))=(9,54,6,324)or(9,72,4,288).(r,a,b,z(G))=(9,54,6,324)\quad\text{or}\quad(9,72,4,288).

Apply Corollary 4.7. Either it supplies a new selected centralizer of index at most eighteen, giving z(G)≤18J(10)=162z(G) \le18J(10) = 162, or it gives z(G)∣54z(G) \mid54 or z(G)∣432z(G) \mid432. Both values 324 and 288 exceed 162, and neither divides 54 or 432. This settles cap ten and completes every cap through sixteen.

Step 6. Induction for all larger even caps. Proceed by strong induction on the even caps k≥18k \ge18. In case (B), t≤k/4<kt \le k/4 < k; the profile bound needed in Step 1 follows from smaller caps. If tt is odd and t≥7t \ge7, use the already proved even cap t+1≤k−2t + 1 \le k - 2; otherwise the bound at tt is already available. Since J(k)≤3(k−4)/3J(k) \le3^{(k-4)/3}, we obtain

z(G)≤T(k):=(k−1)(k−2)3(k−4)/3.z(G) \le T(k) := (k - 1)(k - 2)3^{(k-4)/3}.

The same envelope covers case (A): z(G)≤kP(k−1)≤k3(k−1)/3z(G) \le kP(k - 1) \le k3^{(k-1)/3}, and 3k≤(k−1)(k−2)3k \le(k - 1)(k - 2). At the first cap,

T(18)=272⋅81⋅32/3<272⋅81⋅94=49572<65536=L18.T(18)=272\cdot81\cdot3^{2/3}<272\cdot81\cdot\frac{9}{4}=49572<65536=L_{18}.

The rational factor in the following ratio decreases with kk, so for every even k≥18k \ge18,

T(k+2)T(k)=k(k+1)(k−1)(k−2)32/3<17113694<3<4.\frac{T(k+2)}{T(k)}=\frac{k(k+1)}{(k-1)(k-2)}3^{2/3}<\frac{171}{136}\frac{9}{4}<3<4.

Since Lk+2=4LkL_{k+2}=4L_k at the even caps, the envelope remains strictly below the target. This completes the induction and the proof of the exact profile. □\square

Extremizers and the sharp constant

The cap five requires one elementary classification of alternating maps. It accounts for the two different extremal models at central index sixteen.

Lemma 5.1 (Nonsingular alternating maps in dimension four). Let V=F24V = \mathbb{F}_2^4, let W≠0W \ne0 be a finite-dimensional F2\mathbb{F}_2-space, and let β:V×V→W\beta: V \times V \to W be alternating and bilinear, with values spanning WW. Suppose that λ∘β\lambda\circ\beta is nondegenerate for every nonzero λ∈W∗\lambda\in W^*. Up to linear identifications of VV and WW, precisely two possibilities occur:

(a) W=F2W = \mathbb{F}_2 and β\beta is a nondegenerate symplectic form;

(b) V=F42V = \mathbb{F}_4^2, W=F4W = \mathbb{F}_4, and β\beta is the determinant over F4\mathbb{F}_4.

Proof. The scalar forms constitute a subspace SS of the alternating forms on VV, of dimension dim⁡W\dim W, since the values of β\beta span WW. For an alternating matrix C=(cij)C = (c_{ij}) of size four, its Pfaffian is

Q(C)=c12c34+c13c24+c14c23,det⁡C=Q(C)2.Q(C) = c_{12}c_{34} + c_{13}c_{24} + c_{14}c_{23}, \qquad\det C = Q(C)^2.

Consequently QQ takes the value one on every nonzero element of SS. If SS contained a three-dimensional subspace, the sum of QQ over that subspace would be one in F2\mathbb{F}_2, since there are seven nonzero vectors. But a polynomial of degree at most two on F23\mathbb{F}_2^3 sums to zero: each of its monomials is one at an even number of vectors. Thus dim⁡W≤2\dim W \le2. Dimension one gives (a).

In dimension two, write β=(b,c)\beta= (b,c), where bb, cc, b+cb+c are nondegenerate alternating forms, and define TT by c(x,y)=b(x,Ty)c(x,y) = b(x,Ty). Choose a symplectic basis for bb, so its matrix is J=(0I2I20)J = \begin{pmatrix} 0 & I_2 \\ I_2 & 0 \end{pmatrix}. Write the matrix of cc as

C=(aJ2BBtdJ2),J2=(0110).C = \begin{pmatrix} aJ_2 & B \\ B^{\mathsf{t}} & dJ_2 \end{pmatrix}, \qquad J_2 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}.

Since T=JCT = JC, the two-by-two Cayley–Hamilton identity gives

T2+tr⁡(B)T+(det⁡B+ad)I4=0.T^2 + \operatorname{tr}(B)T + (\det B + ad)I_4 = 0.

Indeed, the diagonal blocks follow from Cayley–Hamilton for BB, and the off-diagonal blocks use BtJ2+J2B=tr⁡(B)J2B^{\mathsf{t}}J_2 + J_2B = \operatorname{tr}(B)J_2. Both TT and T+I4T + I_4 are invertible, since cc and b+cb+c are nondegenerate. Of the four monic quadratics over F2\mathbb{F}_2, only t2+t+1t^2 + t + 1 can annihilate such an operator. Hence

T2+T+I4=0.T^2 + T + I_4 = 0.

Let F4=F2(τ)\mathbb{F}_4 = \mathbb{F}_2(\tau), where τ2+τ+1=0\tau^2 + \tau+ 1 = 0, and let τ\tau act on VV as TT. The operator TT is self-adjoint for bb, because cc is symmetric. The F4\mathbb{F}_4-valued form β~=b+τc\widetilde{\beta} = b + \tau c is alternating and satisfies

β~(x,Ty)=c(x,y)+τ(b(x,y)+c(x,y))=τβ~(x,y).\widetilde{\beta}(x,Ty) = c(x,y) + \tau(b(x,y) + c(x,y)) = \tau\widetilde{\beta}(x,y).

It is therefore bilinear over F4\mathbb{F}_4 and nondegenerate on the two-dimensional F4\mathbb{F}_4-space VV. A choice of basis makes it the determinant, proving (b). Conversely, both listed maps have the required property. For the determinant, if v≠0v \ne0, the map u↦det⁡(v,u)u \mapsto\det(v,u) is onto F4\mathbb{F}_4, so its composition with any nonzero F2\mathbb{F}_2-linear functional is nonzero. Thus no nonzero vv belongs to the radical of any nonzero scalarization, which proves the required nondegeneracy.

Proof of Theorem 1.2. For k=1,2k = 1,2, the absence of a clique of size three forces GG to be abelian. Now assume k≥3k \ge3. Every nontrivial value M(k)M(k) is a prime power. Hence a maximizer is nilpotent, with only one noncentral Sylow subgroup. For k∈{3,4,6}k \in\{3,4,6\} and p=k−1p = k - 1, the noncyclic central quotient of order p2p^2 is Cp2C_p^2. Thus GG has class two and two lifts x,yx,y of a basis, with the centre, generate GG. Their commutator is nontrivial and satisfies [x,y]p=[xp,y]=1[x,y]^p = [x^p,y] = 1, so it generates G′≅CpG' \cong C_p. Conversely, such a scalar symplectic plane gives z(G)=p2z(G) = p^2 and ω(G)=p+1=k\omega(G) = p + 1 = k.

At k=5k = 5, write G=P×A0G = P \times A_0, where PP is its Sylow 2-subgroup and A0A_0 is central of odd order. Then z(P)=16z(P) = 16 and ω(P)≤5\omega(P) \le5. The binary budget gives ω(P)=5\omega(P) = 5 and forces PP to have class two with P/Z(P)≅F42P/Z(P) \cong\mathbb{F}_4^2. The group P′P' is elementary abelian, since [x,y]2=[x2,y]=1[x,y]^2 = [x^2,y] = 1, so the commutator is an alternating bilinear map β:P/Z(P)×P/Z(P)→P′\beta: P/Z(P) \times P/Z(P) \to P' whose values span P′P'.

For every nonzero scalarization, equality in (9) forces m=h=0m = h = 0, hence R=Z(P)R = Z(P). Thus every scalar form of β\beta is nondegenerate. Lemma 5.1 identifies β\beta with either the scalar symplectic form or the determinant on F42\mathbb{F}_4^2. These are exactly the commutator maps of an extraspecial group of order 32 and of UT⁡3(F4)\operatorname{UT}_3(\mathbb{F}_4). Both models have z=16z = 16 and ω=5\omega= 5, as shown in the introduction. Their derived subgroups have orders two and four, so they are not isoclinic. This proves exhaustivity at cap five.

At k=7,8k = 7, 8, apply Theorem 3.3(iii) to the Sylow 3-subgroup. It gives G′≅C3G' \cong C_3, G/Z(G)≅F34G/Z(G) \cong\mathbb{F}_3^4 and ω(G)=7\omega(G) = 7. The converse is the scalar W(3,3)W(3, 3) calculation.

Finally let k≥9k \ge9, r=⌊(k−1)/2⌋r = \lfloor(k-1)/2 \rfloor and z(G)=Lk=22rz(G) = L_k = 2^{2r}. Theorem 3.5 applies with n=2r≥8n = 2r \ge8. Its second alternative is impossible because 2r+3>k2r + 3 > k. Hence G′≅C2G' \cong C_2 and G/Z(G)≅F22rG/Z(G) \cong\mathbb{F}_2^{2r}, with ω(G)=2r+1\omega(G) = 2r + 1. The converse follows from the binary symplectic clique calculation. In each scalar case a symplectic basis identifies the commutator map with that of the corresponding extraspecial group. Together with the cap-five classification, this proves all the asserted isoclinisms.

Proof of Corollary 1.3. For N=ω(G)≥9N = \omega(G) \ge9, the exact profile gives z(G)≤LN≤2N−1<(81/64)2N−1z(G) \le L_N \le2^{N-1} < (81/64)2^{N-1}. For N≤8N \le8, comparison of the eight values of M(N)/2N−1M(N)/2^{N-1} shows that the maximum is 81/6481/64, reached only at N=7N = 7 and z(G)=81z(G) = 81. Theorem 1.2(iv) gives its precise isoclinism class. The same exact formula yields the limit in (2); the extraspecial binary family proves optimality of the asymptotic base two.

Declarations

Funding. The author received no funding for this work.

Competing interests. The author declares no competing interests.

Data availability. No empirical datasets were generated or analysed during this theoretical study.

Use of AI-assisted tools. ChatGPT (OpenAI) was used as an assistant in checking mathematical proofs and in editing and formatting the LaTeX manuscript. The author takes full responsibility for the mathematical content and the final manuscript.

References

  1. [1]N. Andruskiewitsch and M. Graña, From racks to pointed Hopf algebras, Adv. Math. 178 (2003), no. 2, 177–243. doi:10.1016/S0001-8708(02)00071-3.DOI
  2. [2]A. Azad, S. Fouladi and R. Orfi, Maximal subsets of pairwise non-commuting elements of some finite p-groups, Bull. Iranian Math. Soc. 39 (2013), no. 1, 187–192. https://bims.iranjournals.ir/article_341.html.
  3. [3]G. M. Bergman and H. W. Lenstra, Jr., Subgroups close to normal subgroups, J. Algebra 127 (1989), no. 1, 80–97. doi:10.1016/0021-8693(89)90275-5.DOI
  4. [4]Y. Berkovich, Coverings of finite groups by few proper subgroups, Glas. Mat. Ser. III 45(65) (2010), no. 2, 415–429. doi:10.3336/gm.45.2.09.DOI
  5. [5]E. A. Bertram, Some applications of graph theory to finite groups, Discrete Math. 44 (1983), no. 1, 31–43. doi:10.1016/0012-365X(83)90004-3.DOI
  6. [6]A. E. Brouwer, An inequality in binary vector spaces, Discrete Math. 59 (1986), no. 3, 315–317. doi:10.1016/0012-365X(86)90177-9.DOI
  7. [7]A. E. Brouwer and H. Van Maldeghem, Strongly Regular Graphs, Encyclopedia of Mathematics and its Applications, vol. 182, Cambridge University Press, Cambridge, 2022. doi:10.1017/9781009057226.DOI
  8. [8]R. A. Bryce, V. Fedri and L. Serena, Covering groups with subgroups, Bull. Aust. Math. Soc. 55 (1997), no. 3, 469–476. doi:10.1017/S0004972700034109.
  9. [9]M. Ceria, J. De Beule, F. Pavese and V. Smaldore, On large partial ovoids of symplectic and Hermitian polar spaces, J. Combin. Des. 31 (2023), no. 1, 5–22. doi:10.1002/jcd.21864.DOI
  10. [10]A. Y. M. Chin, On non-commuting sets in an extraspecial p-group, J. Group Theory 8 (2005), no. 2, 189–194. doi:10.1515/jgth.2005.8.2.189.DOI
  11. [11]M. Cimráková and V. Fack, Searching for maximal partial ovoids and spreads in generalized quadrangles, Bull. Belg. Math. Soc. Simon Stevin 12 (2005), no. 5, 697–705. https://biblio.ugent.be/publication/358180.DOI
  12. [12]P. Erdős and E. G. Straus, How abelian is a finite group?, Linear Multilinear Algebra 3 (1976), no. 4, 307–312. doi:10.1080/03081087608817122.DOI
  13. [13]P. Etingof, A. Soloviev and R. M. Guralnick, Indecomposable set-theoretical solutions to the quantum Yang–Baxter equation on a set with a prime number of elements, J. Algebra 242 (2001), no. 2, 709–719. doi:10.1006/jabr.2001.8842.DOI
  14. [14]S. Fouladi and R. Orfi, Maximum size of subsets of pairwise noncommuting elements in finite metacyclic p-groups, Bull. Aust. Math. Soc. 87 (2013), no. 1, 18–23. doi:10.1017/S0004972712000111.DOI
  15. [15]M. Graña, Indecomposable racks of order p², Beitr. Algebra Geom. 45 (2004), no. 2, 665–676. https://emis.de/ft/11132.
  16. [16]R. M. Guralnick and G. R. Robinson, On extensions of the Baer–Suzuki theorem, Israel J. Math. 82 (1993), 281–297. doi:10.1007/BF02808114.DOI
  17. [17]P. Hall, The classification of prime-power groups, J. reine angew. Math. 182 (1940), 130–141. doi:10.1515/crll.1940.182.130.DOI
  18. [18]H. Liu and Y. Wang, On non-commuting sets in certain finite p-groups, Algebra Colloq. 22 (2015), no. 4, 555–560. doi:10.1142/S1005386715000474.DOI
  19. [19]B. H. Neumann, Groups covered by finitely many cosets, Publ. Math. Debrecen 3 (1954), 227–242. doi:10.5486/PMD.1954.3.3-4.07.
  20. [20]B. H. Neumann, A problem of Paul Erdős on groups, J. Austral. Math. Soc. Ser. A 21 (1976), no. 4, 467–472. doi:10.1017/S1446788700019303.DOI
  21. [21]L. Pyber, The number of pairwise non-commuting elements and the index of the centre in a finite group, J. London Math. Soc. (2) 35 (1987), no. 2, 287–295. doi:10.1112/jlms/s2-35.2.287.DOI
  22. [22]M. J. Tomkinson, Groups covered by finitely many cosets or subgroups, Comm. Algebra 15 (1987), no. 4, 845–859. doi:10.1080/00927878708823445.DOI
  23. [23]M. Zarrin, On noncommuting sets and centralisers in infinite groups, Bull. Aust. Math. Soc. 93 (2016), no. 1, 42–46. doi:10.1017/S0004972715000696.DOI

Paper details

vlnc:W_xp8ws5ph
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT

View statement
Looking for related papers…
Contents